Click here to


Are you sure ?

Yes, do it No, cancel

Assessing the Influence of Imaging Parameters and Contouring On Meningioma Local Failure Prediction Models

A Witztum*, E Gennatas , G Valdes , T Solberg , D Raleigh , O Morin , University of California San Francisco, San Francisco, CA


(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: To add robustness to a local failure (LF) prediction model by considering radiomic features that are affected by imaging parameters (IPs).

Methods: Meningioma patients (n=230) with MRI IP information (magnetic field strength (MFS), echo time (TE), relaxation time (TR)) were selected. 169 common radiomic (morphological, statistical and textural) features were extracted from a radiomic target volume (RTV), an expanded RTV, and a contracted RTV. Random forest (RF) models predicting MFS, TE, and TR were built to identify associated features. LF-prediction models were built using a 25-fold-stratified-cross-validation in three ways: i) Pre-selection using independent component analysis (ICA) prior to RF, ii) removing features predictive of IPs prior to RF, and iii) finding feature residuals (from a generalized linear model) prior to RF.

Results: Using all features, the RF models built to predict MFS, TR, and TE had mean balanced accuracies (MBAs) of 0.83, 0.95, and 0.59, respectively. LF was predicted with an RF MBA of 0.63 (all features) and 0.59 (without top 20 features from MFS/TR models), suggesting some noise in the features. Using feature residuals to predict LF in an RF model yielded an MBA of 0.63, while the traditional approach of taking 10 ICA components and then building an RF has an MBA of 0.62. Features most predictive of IPs were primarily statistical or intensity based and most predictive of LF were primarily morphological. This explains the small changes between these results. Expanding (MBA = 0.56) and contracting (MBA = 0.59) the RTV by 1 mm reduced the predictive power of LF.

Conclusion: It is expected that some features will be affected by IPs and it is important to remove these from predictive models. We present two methods to remove these features, with one giving comparable or slightly better MBA than techniques that do not correct for this.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email